loading
loading
AI is making security work faster and cheaper — for attackers and defenders. But finding more vulnerabilities is not the same as reducing risk. Teams need to identify what matters in their environment, get it fixed, and verify that they are protected. We’re launching Devin Security 2.0 to help teams complete that work. It combines frontier-model scanning, prioritization grounded in your code, infrastructure, and threat model, and autonomous remediation that turns findings into pull requests your team can review. October 21, 2026 Last week, the PHP and Node.js projects disclosed remote code execution vulnerabilities, affecting a large share of the web. Within hours, every company running PHP or Node was in the same position: a public exploit path, an attacker population that reads the same advisory, and a patch that has to land across container images, OS packages, CI pipelines, and infrastructure configurations, across dozens of repositories. The PHP2Shell exploit is particularly interesting. Signs of one of the underlying flaws had already been identified in 2024, but maintainers had not found a concrete path to remote code execution. Work on a fix stalled amid compatibility trade-offs, limited maintainer resources, and difficulty agreeing on an approach. Without a demonstrated exploitation path, other issues with clearer impact took priority. Wiz reported that frontier models helped its researchers develop and adapt working exploits across multiple applications in hours. That raises a question for any team with a security backlog — are there more cases of deprioritized flaws that AI now makes practical to exploit? At the same time, AI is adding new findings to that backlog. Teams have to reconsider old judgments while investigating a growing stream of reports. Those reports can be accurate and still overwhelm the people responsible for acting on them. As curl maintainer Daniel Stenberg described: “Our current challenge is much harder: high volume, high quality.” AI can help write the patches, too. But engineers still have to resolve uncertain findings, review changes, and take responsibility for what reaches production. More pull requests simply move the bottleneck. To decide what deserves attention first, reviewers need evidence: who can reach the vulnerable code, what an attacker could accomplish, which controls prevent exploitation, and whether the proposed fix closes the attack path. Devin Security 2.0 Devin Security 2.0 connects discovery, prioritization, and remediation in one workflow. The work starts with a finding and follows it through an investigation: establish the attack path, assess its consequences in the customer’s environment, and carry that evidence into a proposed fix. Scan with frontier cyber models Devin Security can now run managed scans with Claude Mythos and other frontier cyber models. Enterprise security teams can use these capabilities through Devin without needing separate access to the models. For scanning your repos, Devin uses an architecture we call Agentic MapReduce, which scales security investigations across your codebase in a cost-efficient manner: - Devin maps your codebase for its routes, authorization boundaries, data access patterns, and other security-relevant code, against a threat model of your repo. - The selected code is batched into groups and investigated in parallel. - Devin aggregates the results, checks candidate findings, removes duplicates, and connects findings that may form a more serious attack path together. - Higher-priority findings are validated in a sandboxed environment against a running instance of your application. This process is model agnostic, allowing us to select the best model for your investigation needs, and use frontier cyber models for your most critical repositories. Selected findings can be assigned to Devin individually or in bulk. Devin checks them against the current code, groups findings that share a fix, and launches independent remediation sessions in parallel. Each session has its own virtual machine, where Devin can run real services, reproduce the issue, implement a targeted fix, and test the result independently. Devin runs the reproduction test and your existing tests, then opens a PR with the investigation and test results for review. This is what makes parallel remediation possible: independent agents doing the engineering work, not just generating patches. Findings remain open until the PR merges and Devin checks that the change addresses the attack path. A new approach to prioritize findings When an investigation is complete, Devin answers two questions to label a finding: - Can an attacker exploit this in your environment? Is the vulnerable code reachable from an untrusted input? Across which trust boundary? What has to be true for the exploit to work — authentication, a specific configuration, user interaction? Is the code live or dead? - What would the damage mean for your organization? How bad is it if one user can read another user’s data? If a user can escalate to org admin? If someone can use your product without paying? Every company answers differently, and these answers can turn an abstract vulnerability into a business priority. Devin answers these questions by developing a threat model with your team. This is a living document, improved with every scan, that determines how your code lives in the real world. The threat model defines every service, host, network policy, and security group; how code is deployed, what talks to what, where data flows, and where the trust boundaries are. Much of this is readable from code and infrastructure definitions; some of it, like which region a service runs in and whether an internal tool is reachable from the internet, is not, which is why building it is collaborative. The threat model drives how Devin prioritizes findings from its investigations, because even strong models need their initial assessments checked. In Anthropic’s Glasswing review, 90.6% of 1,752 reviewed findings initially rated high or critical were valid vulnerabilities, while only 62.4% were confirmed as high or critical. Finding a real bug and assigning the right severity are separate judgments. Ingest findings from any source The same scanning and remediation workflow can now apply to the findings your team already has from other scanners, bug bounty reports, and existing security backlogs. Findings can also be imported from report files, exports, and connected sources, with their original identifiers and links preserved. You can choose to re-triage those findings with frontier models. Parallel Devin sessions examine the reported issues against the current code, identify duplicates, and check whether a vulnerability remains present or has already been mitigated. This lets teams revisit older reports, including those produced by weaker AI models, and update their priorities with evidence explaining the change. When runtime validation is enabled, selected findings are tested against a running build of your application. Devin records what it attempted and observed, including cases where the result remains inconclusive. Each finding remains traceable to its source as its priority and status change. Supported integrations also sync dismissal and reopening decisions back to the original tool. You can then assign findings individually or in bulk to Devin to implement targeted fixes and open pull requests for your team to review. Not every response starts with a code patch. With configured access, Devin can apply a WAF rule, change a security group, or remove public access to a service while a permanent fix is prepared. For recurring flaws, Devin can also help implement broader changes, such as centralizing authorization checks or parameterizing queries across an application, to address the vulnerability class rather than fix each instance separately. Produce compliance evidence at scale The threat model you develop with Devin is the backbone of its prioritization. It keeps your team and Devin focused on what actually needs attention, and provides the reasoning behind those decisions. This matters more than it sounds because of how compliance works. If you operate under FedRAMP, PCI DSS, or enterprise remediation policies, a high- or critical-severity finding can start a remediation clock. Knowing that a finding presents less risk does not, by itself, change the obligation. You need to document why. Compliance processes already provide ways to account for that context. FedRAMP’s POA&M process allows validated false positives to be closed and risk levels to be adjusted when mitigating factors or compensating controls justify it. But you need evidence that can pass the required review. The threat model and Devin’s runtime validation allow you to do this at scale, especially for a backlog of thousands of findings that you can now import into Devin Security. For every finding Devin relabels, it writes out why, in terms of reachability, authentication, data, and trust boundaries, preserves the original CVSS score alongside Devin’s contextual priority and supporting evidence, and hands it to a human to approve and export to the POA&M or GRC system. Devin does not relabel vulnerabilities listed in CISA’s Known Exploited Vulnerabilities (KEV) catalog. Getting started With a threat model, frontier-scanning, ingestion from advisories, and auto-remediation, we are moving towards a future where a response to a security event can happen without human intervention for the first mile. All the pieces are available to you today in Devin Security. Our engineers will work with you from your first scan to merged PRs. To help teams respond to the new threat landscape, Cognition is committing hundreds of millions of dollars in Devin credits. We’re offering current and future enterprise customers free scans with frontier cyber models. New customers can also receive remediation credits to help fix their most critical vulnerabilities. Contact our team to get started with Devin Security and the credits program. If you run security at a large company, the rest of the world is still asleep on this. You’re not. The question is what you do right now to show your organization what the appropriate level of urgency looks like. One good answer: “We have already scanned every repository with the strongest cyber models available and remediated every critical”. Cognition’s CSO, Alex Stamos, is hosting a conversation on the state of the threat landscape and what to do about it on Monday, October 26 at 10:30am PT. Register here. Every second counts.
document fetched 2026-10-27 02:58